Hierarchical electricity time series prediction with cluster analysis and sparse penalty

Abstract

In big data applications, hierarchical time series prediction is an important element of decision-making and concerns the inherent aggregation consistency, which is maintained by reconciliation methods. The paper proposes a novel multiple alternative clustering time series analysis based hierarchical electricity time series prediction method. Instead of adhering the aggregation consistency passively, we first exploit time series mining to construct a hierarchy, and then apply an optimal reconciliation method to improve the prediction accuracy. In particular, k-means clustering method is employed to cluster time series for many times with different k so as to make a large number of time series clusters (patterns), and then the clusters (patterns) based hierarchies are constructed respectively. With the large number of clusters hierarchies and the original geographical hierarchy, an optimal aggregation consistency reconciliation based prediction approach is proposed. Furthermore, the sparse penalty is adapted in our method for “ideal” clusters selection to improve the prediction performance. Compared with the state-of-the-art methods on real-life datasets, our method achieves the improvement of 11.13% and 24.07% accurate one-step ahead forecasts on electricity load and solar power data respectively.